Model training

Train custom vision models with no infrastructure to manage.

Pick a task and a GPU. Pictograph exports your data, trains a custom model on managed GPUs, and returns the weights.

train.py

from pictograph import Client, TrainingRun

client = Client()

# Freeze the dataset into an export, then train on a managed GPU
client.exports.create(
    dataset_name="road-signs",
    name="v1",
    format="coco",
)
run: TrainingRun = client.training.create(
    dataset_name="road-signs",
    export_name="v1",
    pipeline_type="yolox",
    name="signs-detector",
    config={"epochs": 30},
)
client.training.wait_for_completion(run_id=run.id)

client.models.download(
    name="signs-detector",
    output_path="signs_detector.onnx",
)

How it works

From dataset to trained model

No infrastructure to manage. Pictograph converts your dataset to the right format, runs the job, and returns model weights.

Build a dataset

Annotate images in the editor, auto-annotate with SAM3, or import from V7 or Roboflow.
Pick an architecture and GPU

Choose a vision task and a managed GPU. Pictograph handles the data conversion for you.
Train on managed GPUs

Your job runs on managed GPUs with live progress. No Colab limits, no GPU setup.
Get model weights

Receive trained model weights in your registry. Deploy as an endpoint in one click.

Pick your architecture

Five vision tasks, one trained model

Choose the architecture that fits your task. Every pipeline returns model weights, ready to deploy as an endpoint or run in a workflow.

YOLOX Object detection. Fast and edge-deployable.

RF-DETR Strong detection and instance segmentation with pixel masks.

Segformer Semantic segmentation, pixel-wise class maps.

Classification Whole-image labels.

Train from code

Kick off a run from the SDK or the CLI

Call client.exports.create then client.training.create, and read the run before you spend a GPU hour. Or run it from the command line.

Python SDK

train.py

from pictograph import Client, TrainingRun

client = Client()

# Freeze the dataset into an export, then train on a managed GPU
client.exports.create(
    dataset_name="road-signs",
    name="v1",
    format="coco",
)
run: TrainingRun = client.training.create(
    dataset_name="road-signs",
    export_name="v1",
    pipeline_type="yolox",
    name="signs-detector",
    config={"epochs": 30},
)
client.training.wait_for_completion(run_id=run.id)

client.models.download(
    name="signs-detector",
    output_path="signs_detector.onnx",
)

CLI

cli

# Freeze the dataset into an export, then train on a managed GPU
pictograph exports create road-signs --name v1 --format coco

pictograph train start road-signs v1 \
  --pipeline yolox \
  --gpu a10g \
  --name signs-detector

pictograph models download signs-detector --output signs_detector.onnx

Managed GPUs, transparent pricing. Training runs on managed A10G and A100 GPUs. You pay in transparent USD compute credits priced per GPU-minute, with no infrastructure to provision.

Keep exploring

Auto-annotation Label a dataset fast with SAM3 point, box, and text prompts before you train. Deployments Stand up your trained model as an always-on /predict endpoint in one click. Workflows Run your model over images, video, or a dataset in a visual pipeline.

FAQ

Model training FAQ

What models can I train on Pictograph?

You can train object detection (YOLOX, RF-DETR), instance segmentation (RF-DETR), semantic segmentation (SM-PyTorch), and image classification models. Every model exports to ONNX for portable inference.

How much data do I need to train a model?

A minimum of five annotated images is required for the train, validation, and test split, but more labeled data improves accuracy. Use SAM3 auto-annotation to label a dataset quickly before training.

How is training priced?

Training is billed by GPU time on managed A10G GPUs at the underlying rate plus the platform markup, charged from compute credits. You pay only for the actual training run, with no idle cost.

What happens after training finishes?

The trained model is exported to ONNX and added to your model registry with its evaluation metrics. From there you can run batch inference in a workflow, stand it up as a live deployment, or publish it publicly.

Train your first model free

$5/mo free compute. No payment method required.

Start free Read the docs